Correlation network construction for molecular data based on cross validation framework.

Area of Science:

• Bioinformatics
• Systems Biology
• Computational Biology

Background:

• Correlation networks are crucial for understanding biological data.
• Traditional methods like Pearson and partial correlation rely on statistical significance, which is sensitive to sample size.
• This can lead to inaccurate biological network structures due to under- or over-estimation of interactions.

Purpose of the Study:

• To introduce cvCorNet, a novel cross-validation-based method for constructing more accurate biological correlation networks.
• To address the limitations of sample size dependency in traditional correlation network construction.
• To improve the identification of true biological interactions by controlling false positive edges.

Main Methods:

• cvCorNet partitions data into training and test sets.
• Nodewise regression models are fitted on training data to predict networks on test data.
• Optimal network cutoff is determined by maximizing agreement between predicted and empirically estimated networks on test data.

Main Results:

• cvCorNet demonstrates superior performance in recovering underlying correlation structures across various simulated scenarios.
• The method effectively controls false positive edges, leading to more reliable network inference.
• The approach was successfully applied to real-world glycomics and acute myeloid leukemia datasets.

Conclusions:

• cvCorNet offers a robust and accurate alternative for biological correlation network construction.
• The cross-validation approach mitigates issues related to sample size and statistical significance.
• This method enhances the reliability of uncovering latent biological interactions.